论文部分内容阅读
本文系统地讨论了离散型前向神经网络Madaline的输出对网络参数扰动的敏感性计算.首先,根据神经元Adaline的离散特性,提出了一种离散随机技术方法,推导了Adaline敏感性近似计算公式.该方法从理论上解决了已有的连续随机技术方法无法处理的情况,取消了过强的限制条件,并使得从理论上对Adaline敏感性近似计算进行误差分析成为可能.其次,在Adaline敏感性基础上,针对Madaline的结构特性以及处理层间信息传递面临的高复杂度组合计算问题,提出了一种新的基于贡献度大小的取舍策略来计算Madaline敏感性.该策略在计算精度上优于现有的简单平均方法,并降低了计算复杂度.本文给出的敏感性近似计算公式和算法具有形式简单、计算复杂度低、近似误差小、一般性强等优点,大量的仿真模拟实验验证了公式和算法的准确性和有效性.
This paper systematically discusses the sensitivity of Madaline output to the disturbance of network parameters.Firstly, according to the discrete nature of Adaline, a discrete stochastic technique is proposed, and the approximate formula of Adaline sensitivity is derived The method solves the problem that the existing continuous stochastic method can not handle theoretically, cancels the excessively strong restriction condition, and makes it possible to carry out the error analysis theoretically on the approximate calculation of Adaline sensitivity.Secondly, Aiming at the structural characteristics of Madaline and the computational complexity of dealing with the high complexity of inter-layer information transmission, a new trade-off policy based on contribution degree is proposed to calculate the Madaline sensitivity. This method is superior in computational accuracy In the existing simple averaging method, the computational complexity is reduced.The sensitivity calculation formulas and algorithms presented in this paper have the advantages of simple form, low computational complexity, small approximate error and high generality. A large number of simulation experiments The accuracy and validity of the formulas and algorithms are verified.